Update app.py

app.py

@@ -9,15 +9,15 @@ from functools import lru_cache
 # CONFIG
 # ----------------------------
 JSON_FILE = "form-submissions-1.json"
-# 🚩 FIX: Changed the model ID from an embedding model to a generative LLM.
+# Using a suitable generative LLM (Flan-T5 Large)
 MODEL_ID = "google/flan-t5-large"
 HF_API_TOKEN = os.environ.get("HF_API_TOKEN")
+# NOTE: Keeping these temp files for the filtering step, though output format changes
 FILTERED_CSV = "/tmp/filtered_candidates.csv"
 OUTPUT_FILE = "/tmp/outputs.csv"
 BATCH_SIZE = 50
 
 if not HF_API_TOKEN:
-    # This check is good, but ensure the token is set in your environment (or space secrets)
     raise ValueError("HF_API_TOKEN not found in environment. Add it in Space Secrets.")
 
 CATEGORIES = {
@@ -46,7 +46,16 @@ Candidate JSON: {candidate_str}
 Respond only 'Yes' if suitable, otherwise 'No'.
 """
     headers = {"Authorization": f"Bearer {HF_API_TOKEN}", "Content-Type": "application/json"}
-    payload = {"inputs": prompt}
+    
+    # Adding generation parameters for strict, short output
+    payload = {
+        "inputs": prompt,
+        "parameters": {
+            "max_new_tokens": 10, 
+            "return_full_text": False 
+        }
+    }
+    
     try:
         response = requests.post(
             f"https://api-inference.huggingface.co/models/{MODEL_ID}",
@@ -57,16 +66,13 @@ Respond only 'Yes' if suitable, otherwise 'No'.
         response.raise_for_status()
         result = response.json()
 
-        # Check for API error structure
         if isinstance(result, dict) and "error" in result:
              print(f"LLM API Error: {result.get('error')}")
              return "No"
              
-        # Extract the generated text safely and clean it up
         generated_text = result[0].get("generated_text", "No").strip().lower()
         
-        # Flank-T5 often prepends the prompt or a part of it, so we only need the key decision word
-        # We look for 'yes' or 'no' anywhere in the response and prioritize 'yes' if found.
+        # Robust check for 'yes' (in case the model is not perfectly following instructions)
         if "yes" in generated_text:
             return "Yes"
         elif "no" in generated_text:
@@ -76,6 +82,7 @@ Respond only 'Yes' if suitable, otherwise 'No'.
             
     except Exception as e:
         print("LLM call failed:", e)
+        # In case of API failure, it should not be cached as a negative result (but the lru_cache will cache the 'No')
         return "No"
 
 # ----------------------------
@@ -83,13 +90,11 @@ Respond only 'Yes' if suitable, otherwise 'No'.
 # ----------------------------
 def filter_by_roles(category_name):
     job_titles = CATEGORIES[category_name]
-    # NOTE: Assuming 'form-submissions-1.json' exists in the current directory
     try:
         with open(JSON_FILE, encoding="utf-8") as f:
             data = json.load(f)
     except FileNotFoundError:
-        print(f"Error: JSON file '{JSON_FILE}' not found.")
-        return pd.DataFrame(), None
+        return pd.DataFrame(), f"Error: JSON file '{JSON_FILE}' not found."
         
     filtered = []
 
@@ -97,7 +102,6 @@ def filter_by_roles(category_name):
         work_exps = person.get("work_experiences", [])
         if not work_exps:
             continue
-        # Improved: Check if roleName is not None before calling .lower()
         non_fullstack_roles = [
             exp.get("roleName") for exp in work_exps 
             if exp.get("roleName") and "full stack developer" not in exp.get("roleName").lower()
@@ -105,7 +109,6 @@ def filter_by_roles(category_name):
         if not non_fullstack_roles:
             continue
             
-        # Check for role match in the list of titles
         if any(role in job_titles for role in non_fullstack_roles):
             filtered.append({
                 "Name": person.get("name"),
@@ -117,41 +120,41 @@ def filter_by_roles(category_name):
                 "Salary": person.get("annual_salary_expectation", {}).get("full-time","N/A"),
                 "Category": category_name
             })
+            
     if not filtered:
-        return pd.DataFrame(), None
+        # Return a message instead of the CSV path
+        return pd.DataFrame(), f"No candidates found matching roles for category '{category_name}'."
 
     df = pd.DataFrame(filtered)
     df.to_csv(FILTERED_CSV, index=False)
-    return df, FILTERED_CSV
+    # Return a success message
+    return df, f"{len(df)} candidates filtered by role for category '{category_name}'. Ready for LLM check."
 
 # ----------------------------
-# Step 2: LLM recommendations
+# Step 2: LLM recommendations (Modified to return text)
 # ----------------------------
 def llm_recommendations(category_name):
     job_titles = CATEGORIES[category_name]
 
     if not os.path.exists(FILTERED_CSV):
-        # Re-run the filtering step if the CSV is missing
-        df_filtered, _ = filter_by_roles(category_name)
+        # Rerun filtering to ensure the CSV exists
+        df_filtered, msg = filter_by_roles(category_name)
         if df_filtered.empty:
-            return pd.DataFrame(), None
-    else:
-        df_filtered = pd.read_csv(FILTERED_CSV)
-        df_filtered = df_filtered[df_filtered["Category"] == category_name]
+            return msg # Return the error/no-match message
+
+    df_filtered = pd.read_csv(FILTERED_CSV)
+    df_filtered = df_filtered[df_filtered["Category"] == category_name]
 
     if df_filtered.empty:
-        return pd.DataFrame(), None
+        return f"No filtered candidates found for category '{category_name}'. Run Step 1 first."
 
     recommended = []
-    # Drop N/A values before converting to dict, otherwise json.dumps might fail if they are NaN
     df_filtered_clean = df_filtered.fillna('N/A')
     filtered_candidates = df_filtered_clean.to_dict(orient="records")
 
-    # The batching loop is fine, we will rely on the improved call_llm
     for i in range(0, len(filtered_candidates), BATCH_SIZE):
         batch = filtered_candidates[i:i+BATCH_SIZE]
         for person in batch:
-            # Only send necessary info to save context length and cost
             candidate_info = {
                 "Name": person.get("Name"),
                 "Roles": person.get("Roles"),
@@ -160,18 +163,17 @@ def llm_recommendations(category_name):
             candidate_str = json.dumps(candidate_info)
             response = call_llm(candidate_str, category_name, tuple(job_titles))
             
-            # 🚩 IMPROVEMENT: The call_llm function now returns a clean 'Yes' or 'No'
             if response == "Yes":
                 recommended.append(person)
 
     if not recommended:
-        return pd.DataFrame(), None
+        return f"LLM determined no candidates are suitable for the '{category_name}' category."
 
     df_rec = pd.DataFrame(recommended)
-    # Sort by numeric salary
+    
+    # Sort by numeric salary to get the top 5 with lowest expected salary first
     def parse_salary(s):
         try:
-            # Remove currency symbols, commas, and convert to float
             return float(str(s).replace("$","").replace(",","").replace("N/A", str(float('inf'))))
         except:
             return float('inf')
@@ -179,15 +181,26 @@ def llm_recommendations(category_name):
     df_rec["Salary_sort"] = df_rec["Salary"].apply(parse_salary)
     df_rec = df_rec.sort_values("Salary_sort").drop(columns=["Salary_sort"])
     df_top5 = df_rec.head(5)
-    df_top5.to_csv(OUTPUT_FILE, index=False)
+    
+    # 🚩 NEW: Generate Text Output
+    candidate_names = df_top5["Name"].tolist()
+    
+    if not candidate_names:
+        return f"LLM check passed, but sorting resulted in an empty list (unexpected). No recommendations to display."
 
-    return df_top5, OUTPUT_FILE
+    output_text = f"Top {len(candidate_names)} Recommended Candidates for the '{category_name}' Category:\n\n"
+    
+    for i, name in enumerate(candidate_names):
+        output_text += f"{i+1}. {name}\n"
+    
+    output_text += "\nThese candidates were selected based on relevance by the LLM and sorted by lowest expected salary."
+    
+    return output_text
 
 # ----------------------------
 # Show first 5 raw JSON candidates
 # ----------------------------
 def show_first_candidates():
-    # NOTE: Assuming 'form-submissions-1.json' exists
     try:
         with open(JSON_FILE, encoding="utf-8") as f:
              data = json.load(f)
@@ -198,30 +211,31 @@ def show_first_candidates():
         return pd.DataFrame({"Error": [f"Failed to load JSON: {e}"]})
 
 # ----------------------------
-# Gradio interface
+# Gradio interface (Updated)
 # ----------------------------
 with gr.Blocks() as app:
-    gr.Markdown("### Candidate Recommendation Engine")
+    gr.Markdown("# Candidate Recommendation Engine")
     
     gr.Markdown("#### Raw JSON Preview: First 5 Candidates")
     gr.Dataframe(show_first_candidates(), label="First 5 JSON Entries")
 
     gr.Markdown("---")
-    category_dropdown = gr.Dropdown(list(CATEGORIES.keys()), label="Select Category")
+    category_dropdown = gr.Dropdown(list(CATEGORIES.keys()), label="1. Select Category")
 
     # Step 1: Filter by roles
-    filter_button = gr.Button("1. Filter by Roles")
-    filtered_df = gr.Dataframe(label="Filtered Candidates by Roles")
-    download_filtered = gr.File(label="Download Filtered CSV", file_types=[".csv"])
-    filter_button.click(filter_by_roles, inputs=[category_dropdown], outputs=[filtered_df, download_filtered])
+    filter_button = gr.Button("2. Filter Candidates by Roles")
+    filtered_df = gr.Dataframe(label="Filtered Candidates (Preview)")
+    # 🚩 CHANGE: Display a status message for filtering
+    filter_status = gr.Textbox(label="Filter Status", placeholder="Click 'Filter Candidates by Roles' to start.")
+    filter_button.click(filter_by_roles, inputs=[category_dropdown], outputs=[filtered_df, filter_status])
 
     gr.Markdown("---")
 
     # Step 2: LLM Recommendations
-    llm_button = gr.Button("2. Get LLM Recommendations (Requires Step 1 to run first)")
-    llm_df = gr.Dataframe(label="Top 5 Recommended Candidates")
-    download_llm = gr.File(label="Download Recommendations CSV", file_types=[".csv"])
-    llm_button.click(llm_recommendations, inputs=[category_dropdown], outputs=[llm_df, download_llm])
+    llm_button = gr.Button("3. Get LLM Recommendations (Text Summary)")
+    # 🚩 CHANGE: Output is now a Textbox
+    llm_output_text = gr.Textbox(label="Top Candidate Recommendations Summary", lines=10, placeholder="Click 'Get LLM Recommendations' after Step 2 completes.")
+    llm_button.click(llm_recommendations, inputs=[category_dropdown], outputs=[llm_output_text])
 
 if __name__ == "__main__":
     app.launch()